skills/agents-md/SKILL.md
Create, rewrite, or maintain a repository's AGENTS.md as a short dispatcher for coding agents. Use when asked to write an AGENTS.md from scratch, update stale agent instructions, shrink a bloated AGENTS.md, add repo guidance for AI/coding agents, or document where agents should find source-of-truth docs, invariants, validation commands, and workflow rules. Triggers on: "create AGENTS.md", "write an agents file", "update AGENTS.md", "maintain AGENTS.md", "fix our agent instructions", "make this repo easier for Codex/Claude/agents", "rewrite this bloated AGENTS.md".
npx skillsauth add rarestg/rarestg-skills agents-mdInstall this skill globally with one command. Works with Claude Code, Cursor, and Windsurf.
3 of 9 scanners reported clean
Some scanners were skipped, did not run, or reported a non-clean status. Review each row below.
AGENTS.md is a map, not a manual. Its job is to orient the agent, route it
to the real sources of truth, call out non-negotiable constraints, and explain
how to verify work. It should not try to be the truth itself.
Read top-level docs and structure first: README*, existing AGENTS.md, docs
directories, manifests, CI workflows, test/lint configs, and any active
plan/spec/design-doc folders.
Before adding any line, ask:
If several answers are "no", put it in a deeper repo doc or encode it in
tooling — not in AGENTS.md.
ARCHITECTURE.md, design docs, specs, plans,
generated references, quality/security docs)For larger repos, link outward instead of embedding detail. Route to subsystem or nested agent docs rather than duplicating them. For tiny repos, a single concise file may be enough.
A finished AGENTS.md should let a fresh agent quickly answer:
If any answer is missing, the file is incomplete.
Reduce by default, do not expand. Delete obsolete guidance aggressively. Replace duplicated detail with links to canonical docs.
AGENTS.md, don't duplicateAGENTS.md into the full knowledge base unless the repo is tiny.tools
Break large code changes into small, stacked pull requests using vanilla git and the gh CLI. Auto-trigger when implementing a feature or change that spans multiple logical steps, touches several files, or will exceed ~200 changed lines. Also trigger on "stack PRs", "break this into smaller PRs", "stacked diffs", or "create a PR stack". Do NOT trigger for single-file fixes, small bug fixes, or changes under ~200 lines that are a single logical unit.
development
Export a PR's clean inline review comments, CodeRabbit outside-diff comments, and CodeRabbit nitpicks into local files, then triage review feedback through a stack-aware orchestrator workflow with durable reply drafts for follow-up PRs. Use when given a GitHub pull request URL and asked to work through review comments without relying on noisy raw API blobs.
development
Run reusable Graphify-led architecture analysis for codebases using semantic graphs, optional subagent extraction, graph synthesis, source-search validation, graph-shape review, and follow-up refactor planning. Use when asked to analyze repo architecture, god nodes, surprising edges, topology, module boundaries, or graph-derived cleanup/refactor opportunities.
testing
Run a meaningful coding ticket through a delegated delivery workflow: tighten the ticket, assign one ticket owner, delegate implementation, get an independent review, scan for high-value simplification, validate the result, and return a compact outcome packet. Use when the user wants structured agent execution with clear scope, ownership, and review rather than a single-pass implementation. Triggers on: "delegate this ticket", "use a sub-PM", "run this through worker and reviewer", "own this ticket end to end", "send this for independent review", or "close this ticket out with review". Skip trivial fixes and tasks that are still too vague to delegate.